queue-refilllisted

# /queue-refill — auto-refill the experiment queue Keeps the autonomous agent's forward-looking work queue full. This is the "Prioritization Specialist" pattern from arXiv 2604.13018 adapted to our harness. ## When to run - **Scheduled:** every N hours (via `/schedule` or `/loop`), so the queue is always ≥ target when a GPU frees up. - **Reactive:** when `queue.py refill-needed` exits non-zero. - **Manual:** user asks "what should we try next?" or "refill the queue". ## What the skill does 1. **Read recent context from workflow.db:** - Last 20 `[LEARN]` rules (`learnings` table) - All active `dead_ends` for this project - `best_so_far` view (per metric) - `calibration_scorecard` — which hypotheses were over/under-predicted - `experiments` table — recent nodes, stages, failure_class distribution 2. **Read STATE.md** to understand current research direction. 3. **Brainstorm 3–5 new queue items.** Each item must: - Be a single-sentence testable hypothesis - Include a predicted metric delta (forced quantitative prior) - Not overlap an active `dead_ends` direction (query before proposing) - Prefer building on best-so-far (extend the working path) unless there's evidence the path is saturating - Match the current `stage.py get` suggestion (draft → tune → creative → ablation) 4. **Call `queue.py add` for each.** Example: ```bash python3 .claude/scripts/queue.py add \\ --hypothesis "Matrix token init with log-normal std 0.02